Can hardness be screened? Using Machine–learning for targeting mechanical properties in inorganic materials

A. Mansouri Tehrani ^a, A. O. Oliynyk ^a, M. Parry,^b T. D. Sparks,^b J. Brgoch ^a

aDepartment of Chemistry, University of Houston, Houston, TX, 77204, USA
bDepartment of Materials Science and Engineering, University of Utah, Salt Lake City, UT 84112 USA

Can hardness be screened? Using Machine–learning for targeting mechanical properties in inorganic materials

Mansouri Tehrani ^a, A. O. Oliynyk ^a, M. Parry,^b T. D. Sparks,^b J. Brgoch ^a
amansouritehrani@uh.edu
^aDepartment of Chemistry, University of Houston, Houston, TX, 77204, USA
^bDepartment of Materials Science and Engineering, University of Utah, Salt Lake City, UT 84112 USA

Even though the concept of mechanical hardness in a material is very tangible, its dependence on manifold variables ranging from atomic scales to microstructures renders a clear understanding of its composition-structure-property challenging. To explore novel compositions and phase-spaces for high hardness, it is vital to use a reliable model capable of addressing different chemistries. However, even with such a model, hardness, by definition, is an extrinsic property and is generally measured based on the indentation of a material’s surface. Making things more complex, experimental reports are also not consistent and can be performed using different indentation tests (Vickers, Brinell, etc.), loads, and sample purities. Therefore, the discovery of new hard materials traditionally relies heavily on trial-and-error methods and simple design rules. Nonetheless, attempts have been made to correlate hardness to other intrinsic material’s properties such as bulk and shear modulus. Bulk modulus is reported to have a minor correlation with hardness whereas the shear modulus demonstrates a relatively robust relationship. Therefore, we predicted bulk and shear modulus of ≈70,000 binary and ternary compounds from Pearson’s Crystal Database using Machine Learning. The model was constructed implementing a Support Vector Machine (SVM) algorithm using Materials Project database of elastic constants (more than 3000 DFT entries) as the training set. To evaluate the reliability of our SVM model, we synthesized and performed Vickers microindentation measurement of transition-metal disilicides (TMSi₂, TM = V, Cr, Nb, Mo, Ta, W) and rare-earth disilicides (RESi₂, RE = La, Ce, Dy, Gd, Er) due to their range of predicted mechanical properties. Our results show that the RESi₂ predicted hardness values show great agreement with the experiment. TMSi₂ values demonstrate extreme deviation (≈ 50 %) yet qualitatively following the trend. These deviations highlight that B and G are not sufficient to screen for all inorganic materials. Therefore, additional proxies are required to improve the hardness predictions.